Quants need to adapt too ;-)

Let us start with an assumption:

All you need is good point in time data, ranking systems and strategy books.

Then you read some books, academic papers, you implement what you find, run it for a 100 years, and you will make a killing in the markets.

Yes and no. With price only data, that is o.k., but if you combine price and fundamental data, that is a different story.

Have a look at the most successful traders, they are in general not systematic, they are discretionary traders. In all market wizard books, systematic traders realized the worst performance (which was still > 25% ann., so the worst but very good) relative to the discretionary traders. Yes, most of them have systematic parts in their strategy, but at the end they make their decisions discretionary.

Discretionary traders adapt rapidly —> to the market environment and to the hot themes, and they have an intuition for both.

Three examples for change: Static Value, Data Innovation and the AI Factor

Static Value

My favorite subject: The static value factor.

• “Static value” is something like Price / Book or Price / Earnings —> not forward looking

• “Forward value” would be for example —> CurFYEPSMean/Price —> next year’s earnings estimates / price

• “Real value” would be —> the Quant tries to find out about understated intangible assets or understated long term assets + the question if they will be a positive for future earnings developments

So, why is static value not working very well anymore?

There is a reason —> intangibles

Have a look at the following strategy, the ranking system is static value based —>
https://www.portfolio123.com/port_summary.jsp?portid=1820494

Same strategy, just with a buy rule “Rating("Core: Sentiment") > 90”, so value gets combined with earnings sentiment.

Core Sentiment Ranking System:

When we combine value with earnings momentum (or price momentum, which is a proxy for earnings: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4936136) it works fantastic.

There is a good article about this subject from https://www.osam.com (I love OSAM, so many nuggests!):

https://www.osam.com/Commentary/the-factor-archives-value

“

Travis Fairchild demonstrates how book value is not keeping up with the times and highlights similar points to those in the 1935 [!!!] book above.

“Balance sheets prepared under generally accepted accounting principles (GAAP) are doing an increasingly poor job of reflecting the value of shareholders equity. Recent trends have tended to bias assets well below market value which has led to the increased frequency of negative equity and Veiled Value stocks.

Those tend to fall into three main categories…

1) Understated Intangible Assets: Brand names, human capital, advertising, and research and development (R&D) are rarely represented on the balance sheet

2) Understated Long Term Assets: Assets are often depreciated faster than their useful lives

3) Buybacks and Dividends: When buybacks and dividends exceed net income, they create a decrease in equity which can accelerate distortions”

“

Data Innovation

Another example: actuals data (data that gets collected from press releases):

This (above) is a system (it is exclusive to our investment advisor research group, out of sample about 3 months) based on Core: Sentiment + a market cap factor (as a Quality filter) + ActualGr%ttm(#EPS) + ActualGr%PQ(#EPS) (earnings based on actuals).

We know that earnings are super important, if we get the data faster (via press releases) instead of waiting for the next earnings release, we could use it.

The reason why actuals data works well could be (assumption) —> The CEO talks to the CFO and says “hey, we have a great year, our stock price is down, what can we do about it”. And the CFO answers: “well, let us guide higher via a press release”. FactSet captures the data and gets it into the point-in-time database from Portfolio123.com.

And they adapt with change (I mean 1935 is not yesterday ;-))

The AI Factor

My cup of tea: Yes, but not as a black box!

Sure, one can use AI Factor and brute force based on 365 factors (in the AI world they are called features) or more and let the ML Algo find out what are the most important features to build the ranking system.

But I think it is important to use ML Algos which spit out the feature importance and therefore to understand the economical / behavioral drivers the ML Algo finds.

Top feature importance from 365 Factors (abstract):

AI Factor design choices:

Not to bad, OOS Test from 01/18/20 - 01/22/25, Predictor ExtraTrees II was trained from 2003-01-01 to 2019-07-13.

ExtraTrees II Hyperparams: {"n_estimators": 200, "criterion": "squared_error", "max_depth": 8, "min_samples_split": 2}

Best quants start with a high-level assumption, a common-sense economical / behavioral driver, and then (!) design systems based on it.
And they adapt to change.

All the best and best regards.

Andreas

P.S.: In the AI Factor context: Feature selection and design choices by the quant are still important, but that is another blog post ;-)